Degree

Doctor of Philosophy (PhD)

Department

Division of Computer Science and Engineering

Document Type

Dissertation

Abstract

Flash-based key-value systems are widely deployed in today’s data centers for providing high-speed data processing services. These systems deploy flash-friendly data structures, such as slab and Log Structured Merge(LSM) tree, on flash-based Solid State Drives(SSDs) and provide efficient solutions in caching and storage scenarios. With the rapid evolution of data centers, there appear plenty of challenges and opportunities for future optimizations.

In this dissertation, we focus on understanding and optimizing flash-based key-value systems from the perspective of workloads, software, and hardware as data centers evolve. We first propose an on-line compression scheme, called SlimCache, considering the unique characteristics of key-value workloads, to virtually enlarge the cache space, increase the hit ratio, and improve the cache performance. Furthermore, to appropriately configure increasingly complex modern key-value data systems, which can have more than 50 parameters with additional hardware and system settings, we quantitatively study and compare five multi-objective optimization methods for auto-tuning the performance of an LSM-tree based key-value store in terms of throughput, the 99th percentile tail latency, convergence time, real-time system throughput, and the iteration process, etc. Last but not least, we conduct an in-depth, comprehensive measurement work on flash-optimized key-value stores with recently emerging 3D XPoint SSDs. We reveal several unexpected bottlenecks in the current key-value store design and present three exemplary case studies to showcase the efficacy of removing these bottlenecks with simple methods on 3D XPoint SSDs. Our experimental results show that our proposed solutions significantly outperform traditional methods. Our study also contributes to providing system implications for auto-tuning the key-value system on flash-based SSDs and optimizing it on revolutionary 3D XPoint based SSDs.

Committee Chair

Chen, Feng

Available for download on Wednesday, March 03, 2021

Share

COinS